A COMPARISON OF ESTIMATING FOREST CANOPY HEIGHT
INTEGRATING MULTI-SENSOR DATA SYNERGY
——A CASE STUDY IN MOUNTAIN AREA OF THREE GORGES
Lixin Dong, Bingfang Wu *
Institute of Remote Sensing Applications, Chinese Academy of Sciences, Beijing 100101, P. R. China
Datun Road No. 3, P. O. Box 9718, 100101, Beijing, P. R. China, dlx_water@163.com
KEY WORDS: Forest Canopy Height, Lidar, Multisensor Integration, Three Gorges, Spatial models
ABSTRACT:
Forest canopy height is an important input for ecosystem and highly correlated with aboveground biomass at the landscape scale. In
this paper, we make efforts to extracte the maximum canopy height using GLAS waveform combination with the terrain index in
sloped area where LiDAR data were present. Where LiDAR data were not present, the optical remote sensing data were used to
estimate the canopy height at broad scale regions. we compared four aspatial and spatial methods for estimating canopy height
integrating large footprint Lidar system (GLAS) and Landsat ETM+: ordinary least squares regression, ordinary kriging, cokriging,
and cokriging of regression residuals. The results show that (1) the terrain index will help to extract the forest canopy height over a
range of slopes. Regression modles explained 51.0% and 84.0% of variance for broadleaf and needle forest respectively.(2) some
improvements were achieved by adding additional remote sensing data sets. The integrated models that cokriged regression residuals
were preferable to either the aspatial or spatial models alone.The integrated modeling strategy is most suitable for estimating forest
canopy height at locations unsampled by lidar.
1. INTRODUCTION
Measurements of forest structure are critical for biomass estimation (Sun et al., 2008), biodiversity studies (North et al.1999),
fire modelling (Finney, 1998), carbon stock estimation (Skole
& Tucker, 1993), et,al.. Meantime, forest canopy height is an
important input for ecosystem and is highly correlated with biomass (Lefsky, 2005). Traditionally, these attributes have been
measured in field using handheld equipment, which are timeconsuming and limited in scope to mapping at the landscape
scale (Hyde et al. 2006). For passive optical sensors (Landsat
TM/ETM+), it provide useful structure information in the horizontal plane (Cohen & Spies, 1992), but is difficult to penetrating beyond upper canopy layers (Weishampel et al., 2000).
Full waveform digitizing, large footprint LiDAR provides
highly accurate measurements of forest canopy structure in the
vertical plane (Nilsson, 1996, Lefsky et al. 1999). However,
current lidar sensors have limited coverage in horizontal plane
(Lefsky et al. 2002, Hyde et al. 2006).
In present time, due to no single technology is capable of
provide all broad scale information of vertical structure, there
have been several calls for improving the applicability of
remote sening data through multisensor integration (Hudak et al.
2002), combining information from multiple sensor is a promising efforts to improving the accuracy of canopy height estimation at landscape scales (Slatton et al., 2001, Wulder et al.
2004).
scales by integrating the lidar and Landsat TM/ETM+ at the
lidar sample locations. The basic data from GLAS (maximum
canopy height) and biophysical attributes from Landsat
TM/ETM+ (LAI, Forest cover and vegetation indices) were
used for estimation of canopy height. We compared and tested
four widely used empirical estimation methods: ordinary least
squares (OLS) regression, ordinary kriging (OK), and ordinary
cokriging (OCK), and the mixed model of cokriging of
regression residuals.
Our study area lies in mountain area of Three Gorges, which is
representative of the age and structural classes common in the
region. For forests on level ground, the waveform peaks of
canopy surfaces and underlying ground within the footprint
were easily separate. However, over sloped areas, the extent of
waveform increased which inducing some errors. Harding and
Carabajal (2005) point out that the vertical extent of each
waveform increases as a function of the product of the slope
and the footprint size, and returns from both canopy and ground
surfaces can occur at the same elevation. In order to effectively
extracting the canopy height in the sloped area, a new technique
using Terrain Index algorithms was used for estimation of
canopy height GLAS system in Three Gorges.
2. STUDY AREAS AND DATA PROCESSING
2.1 Sduty Areas
Three Gorges area is a key area of the natural protective regions
in China. It is located in 106º00′~111º50′E, 29º16′~31º25′N,
and covers an area of approximately 5.8×104 km2. It lies in the
lower part of the upper reaches of the Yangtze. To the north is
Lidar-Landsat TM/ETM+ integration has immediate relevance
due to the anticipated launches of the Ice, Cloud, and Land
Elevation Satellite (ICESat) (Hudak et al. 2002). In this study,
our mainly objective was to estimate canopy height at broad
* Corresponding author. Email: wubf@irsa.ac.cn..
379
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
location to canopy height. Terrain Indices was defined as the
range of ground surface elevations within n × n sample
windows applied to DEM at the footprint location (Lefsky,
2005). The following equation was used to estimation the
forest canopy height on the sloped area:
the Daba Mountain and to the south is the Yunnan-Guizhou
Plateau. The areas of river valley and flatland in the reservoir
area account for 4.3%; hilly areas, 21.7%; and mountainous
areas, 74%.
The reservoir area is having a typical subtropical monsoon
climate, which is rainy, humid, and foggy in fall, warm in
winter, hot and dry in summer. The average annual temperature
is about 15ºC~19ºC. The average annual precipitation is
1140mm~1450mm and the multi-year average run-off is
405.6×108 m3 in the local river. The main vegetation there is
Evergreen needleleaf forests, Deciduous needleleaf forests,
Evergreen broadleaf forests, Mixed forests, Brushwood and
Croplands. Due to agricultural development and human
activities, the hilly vegetation and natural vegetation will
gradually be replaced by agricultural crops. The forest cover is
low in the reservoir area, with that in the eastern Sichuan
section being only 16~17% and that in the western Hubei
section being 25~38%. The standing timber structure was
simple, being mostly pure forests, with masson pine occupying
70%, mostly young trees. Statistical data in 2001 shows that the
total population in the reservoir area was 19.621,200, including
14.389,300 agricultural population and 5,231,900 nonagricultural population.
h = b 0 ( w − b1 g + b 2 l )
(Lefsky, 2005) (1)
Where
h is the measured canopy height
w is the waveform extent in meters
g is the terrain index in meters
l is the extent of the leading edge in meters
b0 is the coefficient applied to the waveform, when
corrected for the scaled terrain index
b1 is the coefficient applied to th e terrain index
b2 is the coefficient applied to the leading edge
Then,the equation was fit
algorithm (Craig Markwarddt).
using
Levenberg-Marquardt
2.2 Lidar Data and Processing
2.3 Landsat TM/ETM+
ICESat is a spaceborne, waveform sampling lidar system that
using for measuring and monitoring ice sheet and land
topography as well as cloud, atmospheric, and vegetation
properties. The Geo-science Laser Altimeter System(GLAS)
instrument aboard the ICESat satellite launched on 12 January
2003. GLAS received waveforms record 1064nm wavelength
laser energy as a function of time reflected from an ellipsoidal
footprint. It has 70m spot footprints spaced at 175m intervals
(http://icesat.gsfc.nasa.gov/intro.html). In this study, GLAS
data from L2A(October to November 19,2003), L3A(October,
2004),L3C(February-March,2005),L3D(October to November,
2005) and L3F(May-June,2006) were used.
In this study, five Landsat (Enhanced) Thematic Mapper (TM/
ETM+) scenes (path/row: 125/39,126/39,127/39,127/38 and
128/39) from 2002 served as the primary data source to
estimate several spectral vegetation indices(SVIs). Firstly,
geometric correction and atmospheric correction were
performed using the Image Geometric Correction and ATCOR
modules of the ERDAS image processing software respectively.
These images were then rectified to the Gaussian Kruge
projection (Spheroid: Krasovsky; Central meridian: 111º E;
Central latitude: 0; False easting: 500000 meters; False nothing:
0), and were resampled using the nearest neighbour algorithm
with a pixel size of 30m×30m for all bands. The resultant RMS
(Root Mean Square) error was found to be less than 0.5 pixels.
Then some SVIs were calculated which including EVI, NDVI,
ARVI , MSAVI , SARVI , SAVI and the following vegetation
indices:
GLAS have many products (GLA01-GLA16). GLA01 products
provide the waveforms for each shot. The product GLA14
provides Global Land Surface and Canopy elevation. GLA14
doesn’t contain the waveform, but some parameters derived
from the waveform. Firstly, the GLAS waveforms were
smoothed using Gaussian filters with width similar to the
transmitted laser pulse. The noise level before the signal
beginning and after the signal ending were estimated using a
histogram method (Sun et al.2008). And then the signal
beginning and end were identified by a noise threshold. In this
paper, the threshold was set to the noise plus 4 times the
standard deviation. The waveform extent is defined as the
vertical distance between the first and last elevations at which
the waveform energy exceeds a threshold level (Harding and
Carabajal, 2005). Since the canopy height is related to the
ground surface, not the signal ending, the ground peak in the
waveform was found by comparing a bin’s value with those of
the two neighboring bins. If the distance between the first peak
and the signal ending is greater than the half width of the
transmitted laser pulse, the first significant peak found is the
ground peak (Sun et al.2008). The canopy height is defined as
the distance between the signal beginning and the ground peak.
ρ b 2 − ρ b 5 − ρ b1
ρ b 2 + ρ b 5 + ρ b1
ρ − ρb3 − ρb7
VI2 = b 2
ρb2 + ρb3 + ρb7
(2)
ρ b2 × ρ b4
ρ b7
ρ b2 − ρ b7 − ρ b1
VI4 =
ρ b 2 + ρ b7 + ρ b1
(4)
VI5 =
(6)
VI1 =
VI3 =
Where
ρ b7
ρ b2 + 1
(3)
(5)
ρ bi is the reflectance of band bi. Ecologically relevant structural
attributes such as LAI and forest cover have been estimated from these
SVIs.
Prior research has shown that due to Landsat imagery’s
widespread availability and the grain. extent, and multispectral
features make it suitable for a variety of environment
applications at landscape to regional scales.
The distance between the signal beginning and the ground peak
was extracted by above methods when the surface is flat. Over
sloped area, we used the Terrain Indices at the GLAS footprint
380
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
Where, Z * ( μ ) is the primary variable and λα and μα are the
weights and locations of n neighboring samples respectively.
The kriging
weights was forced to sum to one:
3. ESTIMATION METHODS
Firstly, the canopy height datasets were normalized with a
square root transformation (SQRTHT); afterwards, all estimated
SQRTHT values were backtransformed before comparing to
measured height values (Hudak et al. 2002).
n( μ )
λα ( μ ) = 1
∑
α
(11)
=1
3.1 Ordinary Least Squares(OLS) regression
The OLS multiple regression model takes the general form:
3.3 Ordinary CoKriging (OCK)
n
Z = α + ∑ βi ( X i ) + ε
Cokriging is a multivariate extension of kriging and relies on a
linear model of co-regionalization that exploits not only the
autocorrelation in the primary variable, but also the crosscorrelation between the primary variable and a secondary
variable. Cokriging can be graphically represented by the crosssemivariogram, defined as:
(7)
i =1
Where, Z is the forest canopy height, Xi is the i explanatory
variable (SVIs, LAI, forest cover and X and Y location), β i is
the linear slope coefficient corresponding to Xi,
residual error (Kleinbaum et al.1998).
ε
is the
1 N (h)
∑ (zi ( μα ) − zi ( μα + h) )
2 N ( h ) α =1
× (z j ( μα ) − z j ( μα + h ) )
γ ij ( h ) =
3.2 Ordinary Kriging(OK)
Kriging interpolates the sample data to estimate values at
unsampled locations, based solely on a linear model of
regionalization. The linear model of regionalization essentially
is a weighting function required to krig and can be graphically
represented by a semivariogram. The semivariance γ (h) is the
following equations.
γ ( h) =
1 N (h)
2
∑ ( z ( μ α ) − z ( μ α + h) )
2 N (h) α =1
(12)
Where, γ ij (h) is the cross-semivariance between variables I and
j,
z i and z j is
the data value of variable i and j at locations
μα
and μα + h respectively. The OCK estimator of Z * at
location μ takes the form:
(8)
Z * (μ) =
n1 ( μ )
1 =1
γ (h) is semivariance as a function of lag distance h,
N (h) is the number of pairs of data locations separated by h,
n2 ( μ )
∑ λα (μ )Z (μα ) + α∑ λα (μ )Z (μα )
α
1
1
2 =1
2
(13)
2
Where
and
z
is the data value at locations
μα
The traditional OCK operates under two nonbias constraints:
and μα + h
n1 ( μ )
∑ λα
α
(Goovaerts, 1997). And in this study, the exponential models
was used to simulate the nugget, sill, range and the shape of the
sample semivariogram:
⎡
⎣
γ (h ) = c ⎢1 − exp(
1 =1
n( μ )
λα ( μ )Z ( μα )
∑
α
， n (μ )
∑ λα ( μ ) = 0
2
α 2 =1
(14)
2
(9)
Residuals from the OLS regression models were exported and
imported into ARCGIS software for kriging/cokriging. The
same rules and procedures were followed for modelling the
residuals as for modelling the SQRTHT data.
Where, a is the practical range of the semivariogram, c is sill.
Z * (μ ) =
(μ ) = 1
3.4 Integrated
− 3h ⎤
)
a ⎥⎦
The OK model estimates a value Z * at each location
takes the general form:
1
μ
and
4. RESULTS AND DISCUSSION
In this paper, the needle forest and broadleaf forest were
classified for estimation of the canopy height. The accuracy of
classification was validation by our fieldwork.
(10)
=1
381
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
4.1 Estimation of Plot Maximum Canopy Height
Broadleaf
Regression was used to estimate maximum canopy height as a
function of waveform extent and the 3×3 terrain index. The
regression models are following equations for each forest type.
HNeedle =-1.379+0.702(w+0.0028g+0.104l)
(15)
(16)
HBroadleaf =14.716+0.316(w-0.0127g-3.386l)
F
17.517
2.772
9.762
Sig.
.000
.111
.000
Needle Forest
18
Counts.
14
12
30
14
12
10
8
6
4
needle
Linear
2
0
2
4
6
8
10 12 14 16 18 20 22 24
Field Measured Height
24
Field Measured C anopy Height(m)
Field Measured C anopy Height(m)
N eedle F orest
T errain index m odel
R S q Linear = 0.675
22
20
18
14
16
18
20
22
24
26
28
20
15
10
5
Broadleaf
Linear
0
2
4
6
8
10 12 14 16 18 20 22 24
Field Measured Height
20
In order to compare the difference of the results between two
forest type, we have modeled the results with the linear model.
For the needle forest, the R2 coefficient is 69.2%, which greater
than that of the broadleaf forest(50.62%).The regression model
results preserved actual vegetation pattern, but underestimated
taller canopies and overestimated shorter canopies.
18
16
14
16
18
20
22
24
26
Model Predicted C anopy Height(m)
Model Predicted C anopy Height(m)
25
Figure 2. Measured canopy height Vs. predicted canopy height
R S q Linear = 0.673
12
16
R2 = 0.5062
B roadleaf F orest
T errain index m odel
22
59
0
0
24
.000
Broadleaf Forest
30
R2 = 0.692
16
Table 1. Regressions relating Waveform Extent and Terrain
Index to field measured maximum canopy height
26
4.973
The mape of forest canopy height based on the OLS model is
in the Figure 6(a). The predicted canopy height of OLS was
compared with the measured canopy height in field(Fig.2).
Model Predicted Height
Std.
2.635
4.264
3.913
3.053
Table 2. Test of OLS Regression Equations
Model Predicted Height
R2
.840
.510
.530
Forest type
Needle
Broadleaf
All
.634
4.3 Results of OK/COK
Figure 1. Observed maximum canopy height Vs. estimates of
the same, for needle and broadleaf forest
Firstly, the canopy height data were checked(Fig.3). From the
following figure, we can find that the data is submit to the
normal distribution. Hence, we performed the OK and OCK
model in the ARCGIS software. The result of OK and OCK
models are illustrated in Figure4.
When all forest sites were considered in a single regression, the
resulting equation explained 53.0% of variance with an Std. of
3.913. However, individual sites had clear biases. Regression
equations explained between 51.0% and 84.0% of variance for
each forest type(Fig.1 and Tabl.1). Through comparing
observed maximum canopy height and estimation of the same,
the R2 coefficient of needle forest is 67.5%, which greater than
that of the broadleaf forest(67.3%).
4.2 Results of OLS
In this study, the multiple regression models were developed for
needle forest and broadleaf forest respectively. The sample size
of Landsat TM/ETM+ is 60m which similar to the GLAS
footprint size. The models are following regression equation.
(a) Data Histogram
(b) QQPlot
Figure 3. normal distribution of canopy height data
Hneedle =39.118-7.0E-007X(GK) -1E-005 Y(GK)+2.65LAI+
48.482ARVI-0.001EVI+1.532MSAVI+0.032NDVI-20.311
SARVI+8.771SAVI+35.685VI1+5.619VI2+0.029VI3-15.311
VI4+5.701VI5 -3.85FC
(17)
Hbroadleaf =-81.368-2E-007X(GK)+5.7E-005Y(GK)+2.875LAI
+125.038ARVI+0.001EVI-263.005MSAVI+0.083NDVI113.512SARVI+151.085SAVI-11.133VI1+157.214VI20.205VI3-52.567VI4-22.557VI5+1.258FC
(18)
Where, FC is the forest cover, X(GK) and Y(GK) is the x y
location in Gaussian Kruge projection. These models explained
between 55.8%-63.4% of variance at the study area (Table 2).
Forest type
Needle
R2
.558
Std.
3.323
F
5.556
Sig.
.000
Counts.
82
（a）Kriging
382
（b）CoKriging
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
2
R = 0.4738
needle
Linear
Model Predicted Height
0
24
22
20
18
16
14
12
10
8
6
4
2
0
2
Model Predicted Height
24
22
20
18
16
14
12
10
8
6
4
2
0
Needle Forest + CoKriging
6 8 10 12 14 16 18 20 22 24
Field Measured Height
Broadleaf Forest + Kriging
broadleaf
Linear
2
4
6
R2 = 0.4923
2
4
6 8 10 12 14 16 18 20 22 24
Field Measured Height
24
22
20
18
16
14
12
10
8
6
4
2
0
R 2 = 0.6095
broadleaf
Linear
0
2
4
6
8 10 12 14 16 18 20 22 24
Field Measured Height
4.5 Discussion
needle
Linear
24
22
20
18
16
14
12
10
8
6
4
2
0
8 10 12 14 16 18 20 22 24
needle
线性 (needle)
0
2
4
6
The results from this study confirm that forest height can be
estimated using GLAS waveform combination with the terrain
index in sloped area. Regression equations explained 51.0% and
84.0% of variance for broadleaf and needle forest respectively.
the result of this work indicate that the terrain index will help to
extract the forest canopy height over a range of slopes.
8 10 12 14 16 18 20 22 24
Field Measured Height
Broadleaf Forest + CoKriging
2
R = 0.5955
Integration of GLAS and Landsat TM/ETM+ data using
empirical modeling procedures can be used to improve the
utility of both datasets for forestry applications. In this study,
four integration techniques: OLS,OK,OCK and OCK+OLS
models, were compared. In total, the integrated technique of
ordinary cokriging of the height residuals from an OLS
regression model proved the best method for estimating the
forest canopy height. In future work, to improve the accuracy of
the cnaopy height estimations and test the integration models in
the sloped area once lidar sample data become readily available.
broadleaf
Linear
0
Field Measured Height
R 2 = 0.6444
Figure 7. Measured canopy height Vs. predicted canopy height
0
2
R = 0.5008
0
24
22
20
18
16
14
12
10
8
6
4
2
0
4
Model Predicted Height
Model Predicted Height
Needle Forest + Kriging
Broadleaf Forest+CoKriging+Regression
Needele Forest+CoKriging+Regression
24
22
20
18
16
14
12
10
8
6
4
2
0
Model Predicted Height
From figure4 and figure5, we can find that the result of OK is
similar to that of OCK model. But comparing the predicted
canopy height with the field measurement of the canopy height,
the precision of OCK model is prior to that of the OK. However,
the R2 coefficient of needle forest is less than that of the
broadleaf forest. Cokriging proved slightly more accurate than
kriging. The spatial models, kriging and cokriging, produced
greater biased results than regression and poorly reproduced
vegetation pattern. This mar be related with the distribution of
lidar sampling points.
OK/OCK. The the R2 coefficient of broadleaf forest is also
greater than that of the OLS (50.62%). However, the R2
coefficient of needle forest is less than that of the OLS(69.2%).
This is related with the distribution of lidar sampling points. An
equitable distribution of lidar sampling points proved critical
for efficient lidar_Landsat TM/ETM+ integration (Hudak et al.
2002).
Model Predicted Height
Figure 4. canopy height of OK and OCK models
2
4
6
8 10 12 14 16 18 20 22 24
Field Measured Height
Figure 5. Compare of the results of OK and OCK models
4.4 Results of integration model
Due to the precision of OCK is greater than that of OK. Here,
we have only discussed the integration model of
‘OCK+regression’. Residuals from the OLS regression were
imported into cokriging. The result was illustrated in the
figure6(b).
REFERENCES
Craig Markwarddt,
idl.html.
http://cow.physics.wisc.edu/~craigm/idl/
Cohen, W.B., Spies, T.A., 1992. Estimating structure attributes
of Douglas-fir/western hemlock forest stands from Landsat and
SPOT imagery. Remote Sensing of Environment, 41, pp.1-17.
Finney, M.A., 1998. FARSITE: Fire area simulator –model
development and evaluation. (RP-4):1-+.
Hudak, A.T., Lefsky, M.A., Cohen, W.B., Berterretche, M.,
2002. Integration of lidar and Landsat ETM+ for estimating
and mapping forest canopy height. Remote Sensing of
Environment, 82, pp.397-416.
（a）Regression
Hyde, P., Dubayah, R., Walker, W., Blair, J.B., Hofton, M.,
Hunsaker, C., 2006. Mapping forest structure for wildlife
habitat analysis using multi-sensor(LiDAR,SAR/INSAR,ETM+,
Quickbird) synergy. Remote Sensing of Environment, 102,
pp.63-73.
（b）CoKriging+Regression
Figure 6. canopy height of integration model
Through comparing the field measurement of canopy height
with the predicted canopy height, the R2 coefficient of needle
forest is 64.44%, which greater than that of the broadleaf
forest(60.95%). Obviously, this results are greater than that of
Goovaerts P., 1997, Geostatistics for natural resources
evaluation. New York: Oxford University Press, pp.483-484.
383
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
surface topography and vegetation height. IEEE Transactions
on Geoscience and Remote Sensing, 37(5), pp.2620-2624.
Kleinbaum, D.G., Kupper, L.L.,Muller, K.E., Nizam, A.,1998.
Applied regression analysis and other multivariable methods
(3rd ed.), Pacific Grove, CA: Duxbury Press.
Sun, G., Ranson, K.J., Masek, A., Fu, A., Wang, D., 2008.
Predicting tree height and biomass from GLAS data. Science,
260, pp.1905-1909.
Lefsky, M.A., Harding D.J., Keller, M., Cohen, W.B., 2005.
Estimations of forest canopy height and aboveground biomass
using ICESAT. Geophysical Research Letters,32, L22s02,
doi:10.1029/2005GL0239712.
Lefsky, M.A., Harding, D., Cohen, W.B., Parker, G.G., Shugart,
H.H., 1999. Surface lidar remote sensing of basal area and
biomass in deciduous forests of eastern Maryland, USA.
Remote Sensing of Environment, 67, pp.83-98.
North,M.P., Franklin, J.F., Carey, A.B., Forsman, E.D., Hamer,
T., 1999. Forest stand structure of the northern spotted owl’s
foraging habitat. Forest Science, 45(4), pp.520-527.
Weishampel, J.F., Blair, J.B., Knox, R.G., Dubayah, R., Clark,
D.B., 2000. Volumetric lidar return patterns from an oldgrowth tropical rainforest canopy. International journal of
Remote Sensing, 21(2), pp.409-415.
Wulder, M., Hall, R., Coops, N.C., Franklin, S., 2004. High
spatial resolution remoterly sensed data for ecosystem
characterization. Bioscience, 54(6), pp.511-521.
http://icesat.gsfc.nasa.gov/intro.html
Nilsson, M., 1996. Estimation of tree heights and stand volume
using an airborne lidar system, Remote Sensing of
Environment, 56,pp.1–7.
ACKNOWLEDGEMENTS
This work was supported by Three Gorges Project Construction
Committee of the State Council Project ‘The Dynamic and
Real-time Monitoring of Ecology and Environment in Three
Gorges Project’ (SX 2002-004). The authors thank Distributed
Active Archive Center of National Snow and Ice Data Center
(NSIDC) for providing the GLAS data and many helps.
Skole, D., Tucker, C., 1993. Tropical deforestation and habitat
fragmentation in the Amazon: Satellite data from 1978 to 1988.
Science, 260, pp.1905-1909.
Slatton, K.C., Crawford, M.M., Evans, B.L., 2001. Fusing
interferometrc radar and laser altimeter data to estimate
384